← All projects
Supporting case study · Machine Learning

Optimizing a Regression Tree for Used-Car Pricing

Diagnosed overfitting and tested 216 parameter combinations to improve out-of-sample used-car price predictions.

RapidMinerRegression TreesCross-ValidationHyperparameter Tuning
Optimizing a Regression Tree for Used-Car Pricing project graphic
216parameter combinations
5-foldcross-validation
1,309baseline RMSE
1,255optimized RMSE
The challenge

What needed to be understood.

The unrestricted tree fit the training data extremely well but generalized poorly. The task required diagnosing overfitting and selecting parameters using validation rather than training fit.

Approach

How the analysis was built.

  1. Built a full regression tree and evaluated its holdout performance.
  2. Tested combinations of maximum depth, minimum leaf size, and minimum split size.
  3. Used five-fold cross-validation across 216 parameter combinations.
  4. Compared variable importance and predictive error before and after tuning.
Key findings

What the evidence showed.

  • Vehicle age, kilometers driven, quarterly tax, and automatic air conditioning were the most influential specifications.
  • The unrestricted tree overfit the training data.
  • The selected configuration reduced holdout RMSE from roughly 1,309 to 1,255.
Business recommendations

What should happen next.

  • Select tree complexity with cross-validation rather than visual fit to training data.
  • Monitor performance by vehicle segment and price range.
  • Add condition, service history, location, and market-demand variables for stronger pricing accuracy.
Limitations and integrity

What this project does—and does not—prove.

The selected dataset contains a limited set of vehicle characteristics and represents an academic prediction exercise rather than a production appraisal system.

Continue exploring

More projects, methods, and business questions.